flowjax 4.0.2

Normalizing flow implementations in jax.

flowjax

---

Normalising flows in JAX. Training a flow can be done in a few lines of code:

from flowjax.flows import BlockNeuralAutoregressiveFlow
from flowjax.train_utils import train_flow
from flowjax.distributions import Normal
from jax import random

data_key, flow_key, train_key = random.split(random.PRNGKey(0), 3)

x = random.uniform(data_key, (10000, 3))  # Toy data
base_dist = Normal(3)
flow = BlockNeuralAutoregressiveFlow(flow_key, base_dist)
flow, losses = train_flow(train_key, flow, x, learning_rate=0.05)

# We can now evaluate the log-probability of arbitrary points
flow.log_prob(x)

The package currently supports the following:

• Supports both CouplingFlow (Dinh et al., 2017) and MaskedAutoregressiveFlow (Papamakarios et al., 2017) architectures
• Supports common transformers, such as AffineTransformer and RationalQuadraticSplineTransformer (the latter used in neural spline flows; Durkan et al., 2019)
• BlockNeuralAutoregressiveFlow, as introduced by De Cao et al., 2019

For more detailed examples, see examples.

Installation

pip install flowjax

Warning

This package is new and may have substantial breaking changes between major releases.

TODO

A few limitations / things that could be worth including in the future:

• Support embedding networks (for dimensionality reduction of conditioning variables)
• Add batch/layer normalisation to neural networks
• Training script for variational inference
• Add documentation

Related

We make use of the Equinox package, which facilitates object-oriented programming with Jax.

Authors

flowjax was written by Daniel Ward <[e-mail]>.